Replit's New Agent 3
A browser-driven test harness in a controlled environment, not a bigger model, is what lets the agent iterate autonomously for hours.
Replit's Agent 3 is impressive in early testing. The advance that matters is not the model but the harness around it: by controlling the editor, preview, auth, and deploys, Replit lets the agent click through a UI, check that it worked, and keep fixing until it passes, for hours at a time.
My early tests with this are very impressive.
What actually changed?
The interesting advance here is not a bigger model; it is embedding a browser-driven test harness the agent trusts and iterates against. By controlling the environment (editor, preview, auth, deploys), Replit can let the agent click through UIs, assert behaviors, and refactor until green, then keep going for hours under "Max Autonomy." That's cheaper and faster than generic desktop automation precisely because it narrows the domain.
Why the narrow domain wins
- A controlled environment: editor, preview, auth, and deploys
- The agent can check its own work in a real browser
- It iterates until tests pass, then keeps going
- A narrow domain is cheaper and faster than generic desktop automation